Prompt-to-Paper: Automated Manuscript Generation

• The paper introduces a multi-agent system that transforms prompts into scholarly papers by iteratively optimizing outputs and quality scores.
• Prompt-to-Paper is a research direction that establishes a formal prompt engineering framework, incorporating chain-of-thought and reflective self-revision techniques.
• The automated bioinformatics pipeline integrates retrieval-augmented generation, an autonomous coding agent, and an eight-dimensional quality scorer to ensure factuality and reproducibility.

Searching arXiv for the specified papers and closely related work to ground the article. Prompt-to-Paper denotes a research direction and, in one usage, a concrete multi-agent system for transforming prompts into structured scholarly outputs. In the broader prompt-engineering literature, the phrase is associated with the progression from prompt specification to rigorous, utility-driven generation, grounded in a formal view of prompt optimization over the conditional distribution $P(r \mid p)$ (Amatriain, 2024). In a narrower and more recent bioinformatics setting, Prompt-to-Paper names a fully automated pipeline, designated “RLEv4,” that combines deterministic retrieval-augmented generation, an autonomous coding agent, and an eight-dimensional quality scorer inside a context-rich iterative improvement loop with scheduled deep research cycles (Kamran et al., 5 Jul 2026). Together, these formulations position Prompt-to-Paper at the intersection of prompt engineering, agentic workflows, retrieval grounding, executable experimentation, and manuscript evaluation.

1. Conceptual basis in prompt engineering

Prompt-to-Paper rests on the formal framework of prompt engineering in which a pre-trained LLM defines a conditional distribution $P(r \mid p)$ over responses $r$ given prompts $p$, and prompt engineering seeks a prompt $p^*$ that maximizes an expected utility over generated responses (Amatriain, 2024). The formulation given is

$$p^* = \arg\max_p \mathbb E_{r\sim P(\cdot \mid p)} [U(r)] ,$$

with common utility functions including task accuracy, factuality score, and style or sentiment match. A second formulation treats prompt optimization as maximizing the log-likelihood of a target output $r^*$ when such an output is known:

$p^* = \arg\max_p \log P(r^* \mid p) .$

Because the prompt is itself text, the account explicitly states that direct gradient-based optimization is not possible; manual engineering, heuristic search, and automated prompt-engineering methods are therefore used instead (Amatriain, 2024).

This formalization is significant because it frames Prompt-to-Paper not merely as a writing aid, but as a controlled optimization problem over downstream manuscript quality. In that sense, Prompt-to-Paper can be understood as an instantiation of prompt engineering in which the utility function is expanded from local response quality to document-level criteria such as factuality, structural completeness, grounding, and reproducibility. This suggests that “paper generation” is treated not as a single-shot decoding problem but as an iterative search process over prompt-conditioned outputs.

2. Prompt construction workflows and advanced prompting methods

The underlying prompt-design workflow described for prompt engineering consists of seven steps: task definition; template and skeleton design; in-context examples and their order; formatting and special tokens; tone and instruction style; parameter tuning; and iterative testing and refinement (Amatriain, 2024). The workflow specifies explicit output formats such as JSON, bullet list, or essay; distinguishes zero-shot, few-shot, and chain-of-thought prompting; and recommends a template that separates instructions, examples, and user inputs. It also notes that, in few-shot settings, diverse examples should cover the input–output space, and that placing instructions before examples and ordering examples from simplest to most complex often boosts performance.

Chain-of-Thought (CoT) prompting and Reflection are presented as advanced methods directly relevant to Prompt-to-Paper. CoT augments the model output with an intermediate reasoning chain $c$, yielding the decomposition

$$P(r \mid p) = \sum_{c} P(c, r \mid p) = \sum_{c} P(c \mid p)\,P(r \mid p, c),$$

and is operationalized through templates such as “Let’s think step by step” (Amatriain, 2024). Reflection adds a self-revision phase in which an initial answer $P(r \mid p)$0 is critiqued via a generated message $P(r \mid p)$1, followed by a revised answer $P(r \mid p)$2. The formulation given is

$P(r \mid p)$3

The text states that this two-stage process often yields more accurate, coherent outputs because self-critique identifies and corrects hallucinations or mis-reasoning.

These methods are directly aligned with Prompt-to-Paper’s objective of producing manuscript text that is not only fluent but also revisable under explicit quality criteria. A plausible implication is that Prompt-to-Paper operationalizes CoT and Reflection at the document level, where intermediate reasoning, critique, and revision are distributed across sections, agents, and evaluation loops rather than confined to a single answer string.

3. Agentic formulation and system architecture

The more specific Prompt-to-Paper system in bioinformatics is described as a fully automated, multi-agent pipeline with three core modules: deterministic retrieval-augmented generation (RAG), an autonomous coding agent, and an eight-dimensional automated quality scorer (Kamran et al., 5 Jul 2026). These modules are integrated by a context-rich iterative improvement loop with scheduled deep research cycles. The system includes four principal roles: a Leader Agent, identified as deepseek-v4-pro; a Worker Agent, identified as deepseek-chat; a Coding Agent; and a Scoring Module composed of a hybrid G-Eval judge, corpus metrics, heuristic features, and a hallucination auditor.

The agentic architecture inherits from the general LLM-based agent pattern in which an autonomous loop combines perception, reasoning or planning, tool invocation, action execution, and state update (Amatriain, 2024). The general pseudocode builds prompts from an internal state $P(r \mid p)$4, samples a thought-action pair from the LLM, invokes tools when required, and updates the state with tool outputs or feedback. The same source identifies several prompt-engineering patterns for such agents: ReAct, which interleaves reasoning with explicit “Action:” and “Observation:” markers; ReWOO, which performs meta-planning before observations; and DERA, which uses multi-agent dialogue among specialized roles (Amatriain, 2024).

In the Prompt-to-Paper instantiation, the Leader Agent plans high-level JSON-based pipeline architectures, judges quality, performs gap finding, and conducts re-manuscripting on deep cycles; the Worker Agent drafts sections, produces code snippets, and handles JSON planning subtasks; the Coding Agent generates, executes, and validates Python scripts for real bioinformatics experiments; and the Scoring Module produces the quality assessment (Kamran et al., 5 Jul 2026). Internal state management is described as being handled by the surrounding application, which concatenates past interactions up to a token limit into the next prompt, while memory modules or databases can store longer-term facts and preferences (Amatriain, 2024).

4. Deterministic grounding and executable experimentation

A central distinction of Prompt-to-Paper in the bioinformatics system is its emphasis on deterministic grounding and genuine execution. The abstract states that existing automated manuscript-generation systems suffer from three deficiencies: generated claims are not deterministically grounded in verifiable literature, experimental results are frequently fabricated rather than executed, and no standardized, multi-dimensional framework exists for assessing publication-level quality and rigor (Kamran et al., 5 Jul 2026). Prompt-to-Paper addresses these issues through a deterministic retrieval-augmented generation pipeline with section-aware relevance scoring and snowball citation expansion, an autonomous coding agent that executes real computational biology experiments, and a quality scorer with hallucination penalties.

The retrieval component begins from a corpus of 60–100 papers obtained via Semantic Scholar or Tavily API and scored section-wise using SPECTER2, with BM25+ as fallback (Kamran et al., 5 Jul 2026). For a topic $P(r \mid p)$5 and candidate paper $P(r \mid p)$6 with sections $P(r \mid p)$7, the relevance score is

$P(r \mid p)$8

Starting from up to 60 top-ranked seed papers, snowball citation expansion is performed for two iterations using references and citations from Semantic Scholar, retaining candidates whose relevance satisfies $P(r \mid p)$9 for $r$0, and stopping at a corpus size of 100 (Kamran et al., 5 Jul 2026). The final corpus is indexed in ChromaDB using precomputed document embeddings.

The autonomous coding agent receives a bioinformatics task and a JSON plan of algorithmic steps, then iteratively generates Python code, checks syntax with ast.parse, executes code in a sandbox with a one-hour timeout, and extracts results.json and plot.png upon success (Kamran et al., 5 Jul 2026). The environment is Python 3.10 in Docker with Biopython, NumPy, Matplotlib, and a rigor.py module supporting permutation tests, bootstrap confidence intervals, Benjamini–Hochberg correction, and Cohen’s $r$1. The report gives a concrete case-study result for substitution matrix eigenspectrum computation: first eigenvalue $r$2, permutation $r$3-value $r$4, and 95% bootstrap confidence interval for $r$5 equal to $r$6 (Kamran et al., 5 Jul 2026).

The system further includes pre-execution syntax gating, five-retry LLM repair for syntax errors, execution timeout and maximum-attempt guards, and a numeric-mismatch audit against canonical results.json (Kamran et al., 5 Jul 2026). These mechanisms indicate that Prompt-to-Paper is designed to replace synthetic outputs with genuine numerical results rather than merely simulate computational experiments. This suggests a methodological shift from language-only manuscript synthesis toward tool-mediated scientific workflow execution.

5. Quality scoring and iterative manuscript revision

Prompt-to-Paper’s quality-control layer is defined as an eight-dimensional automated quality scorer over Novelty, Contribution, Soundness, Presentation, Reproducibility, Grounding, Gap Relevance, and Structural Completeness (Kamran et al., 5 Jul 2026). The scorer is described as a three-tier hybrid blend: Tier 1 is an LLM judge with weight 55%; Tier 2 consists of corpus metrics with weight 25%; Tier 3 uses heuristic features with weight 20%. The overall score $r$7 lies in $r$8.

The per-dimension composite score is

$r$9

Tier 1 begins from raw judge outputs $p$0, z-normalized and clipped to produce

$p$1

Tier 2 uses corpus metrics, with key-term coverage $p$2 and unique-cite ratio $p$3:

$p$4

except that grounding uses $p$5. Tier 3 normalizes surface features such as word count, table count, math density, and section completeness to $p$6 (Kamran et al., 5 Jul 2026).

Hallucination is handled through a deterministic penalty. Let $p$7 be the fraction of citation markers outside the corpus range, $p$8 the count of numeric mismatches versus canonical results.json, and $p$9 citation coverage. Then

$p^*$0

The overall quality score is

$p^*$1

An equivalent notation is also given:

$p^*$2

Quality scoring drives the revision loop. The improvement algorithm maintains the best manuscript and score, performs prose polishing in most iterations, and triggers deep research cycles at scheduled intervals (Kamran et al., 5 Jul 2026). For prose revisions, it identifies the weakest scoring dimension, routes the manuscript to one of three actions, and accepts a revision if the overall score improves, or if the targeted dimension improves under an overall tie. The three routing actions are add_analysis for Soundness, Contribution, and Reproducibility; gather_evidence for Grounding, Novelty, and Gap Relevance; and rewrite for Presentation and Structural Completeness. On three consecutive failures for the same weakest dimension, the system escalates to the next action in the ladder (Kamran et al., 5 Jul 2026).

Deep research cycles invoke gap finding, code extension, real execution, and full re-manuscripting under a “never-regress” guard that allows a score drop of at most 0.01 (Kamran et al., 5 Jul 2026). With IMPROVE_ITERS = 60, the schedule becomes $p^*$3, placing deep cycles at iterations $p^*$4. In effect, Prompt-to-Paper alternates between local rhetorical refinement and global evidence regeneration.

6. Empirical results, tooling context, and related testing concerns

The Prompt-to-Paper bioinformatics study validates the system on five case studies and reports that all five compiled submission-formatted PDFs with zero out-of-range citations (Kamran et al., 5 Jul 2026). The improvement loop raises manuscript quality by an average of +17.96 points on a 0–100 scale, with a maximum increase of +26.04. The average initial score is 45.60, the average final score is 63.56, average citation coverage is 43.2%, and average cost is 0.309 USD. The five case studies and reported scores are as follows.

Case Study	Init Q	Final Q
Subst. Matrix Eigenspectrum	51.84	68.55
TP53 Hotspot Mutation (DDG-proxy)	48.90	63.20
CpG Island Detection	37.16	63.20
CAI Bootstrap Sensitivity	50.00	61.96
Codon Usage & Shannon Entropy	40.10	60.88

External checks are partial but explicit. Automated $p^*$5, rescaled to a 10-point scale, averages 6.36; three independent LLMs—Claude, GPT-4o, and DeepSeek—average 7.39/10; and a single human bioinformatics reviewer assigns an average of 7.0/10 (Kamran et al., 5 Jul 2026). The system uses approximately 1.8 million tokens per run, with a worker-to-leader ratio of approximately 2.44×, requires 100–190 minutes at 60 iterations, and costs approximately 0.31 USD per paper. The abstract further states that the cost is approximately 0.31 USD per paper and that the system is 5× cheaper than prior LLM agents (Kamran et al., 5 Jul 2026).

Citation fidelity is emphasized: all five runs have zero bad citations, and Recall@20 is approximately 0.93 using a PaSa-style metric against an internal top-20 (Kamran et al., 5 Jul 2026). No formal $p^*$6-values are reported for the overall improvement study, although the report states that the consistent gains of +11.96 to +26.04 points across heterogeneous problems, together with convergence between independent LLM and human reviewer rankings, provide strong empirical evidence of robust improvement.

Within the wider tooling ecosystem of prompt engineering, Prompt-to-Paper sits alongside frameworks such as LangChain, Semantic Kernel, Guidance, Nemo Guardrails, LlamaIndex, FastRAG, Auto-GPT, and AutoGen (Amatriain, 2024). Their reported niches range from template management and chaining to safety constraints, retrieval-augmented generation, and multi-agent application construction. The survey states that real-world applications often combine several of these tools to balance creativity, reliability, and safety. Prompt-to-Paper’s architecture is consistent with that view, because it integrates retrieval, agent coordination, execution, and scoring rather than depending on a single prompt-template abstraction.

A related concern is prompt robustness and regression testing. PromptPex treats prompts as code-like artifacts, extracts input specifications and output rules, and generates rule-based and inverse-rule unit tests to identify invalid outputs and regressions across models (Sharma et al., 7 Mar 2025). Its motivation is that prompt behavior is model-dependent, under-specified, and prone to drift under prompt edits or model upgrades. Although PromptPex is not a manuscript-generation framework, its emphasis on explicit specifications, compliance evaluation, and model-sensitive testing highlights a complementary dimension of Prompt-to-Paper: beyond generating papers, prompt-defined scientific pipelines may also require systematic prompt-level regression suites. This suggests that evaluation of Prompt-to-Paper systems may eventually extend from document quality alone to specification-level robustness under model change.